Transformer imputation in CTG: a length-dependent evaluation of reconstruction methods.

Area of Science:

• Medical Informatics
• Signal Processing
• Artificial Intelligence

Background:

• Intrapartum cardiotocography (CTG) analysis is increasingly reliant on computer and AI methods.
• Fetal heart rate (FHR) signal loss is a common challenge, impacting interpretation accuracy.
• Lack of consensus exists on acceptable missing FHR data lengths for reliable reconstruction.

Purpose of the Study:

• To identify optimal imputation methods for missing FHR data in CTG.
• To define the clinical limits of valid missing-segment lengths for FHR signals.
• To establish a baseline for preprocessing in CTG AI research and clinical use.

Main Methods:

• Utilized an open FHR dataset (CTU-UHB) with artificially introduced data gaps of varying lengths.
• Compared a Transformer-based model against linear and spline interpolation using RMSE and correlation metrics.
• Assessed imputation task difficulty by quantifying similarity between original and removed data.

Main Results:

• Spline interpolation showed significantly worse numerical accuracy (RMSE) compared to other methods.
• The Transformer model generally outperformed linear interpolation in both numerical accuracy and waveform preservation (correlation).
• All imputation methods degraded in performance as gap lengths increased, with reliability notably decreasing beyond 30 seconds.

Conclusions:

• The Transformer model serves as an effective baseline for FHR imputation, balancing waveform and numerical accuracy.
• This study clarifies clinical limits for missing FHR data, indicating reduced reliability for gaps exceeding 30 seconds.
• Findings provide essential guidance for standardizing CTG AI preprocessing and clinical implementation.